Luke Oakden-Rayner

@DrLukeOR

Director of Research @ nRAH Medical Imaging. Researcher @ Australian Institute for Machine Learning. Blogger, radiologist, PhD student. Views my own.

Adelaide, South Australia
Vrijeme pridruživanja: studeni 2016.

Tweetovi

Blokirali ste korisnika/cu @DrLukeOR

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @DrLukeOR

  1. Prikvačeni tweet
    14. lis 2019.

    "Improving Medical AI Safety by Addressing Hidden Stratification" is new research by () and myself (). We argue HS can explain failures of pre-clinical testing, and propose an achievable regulatory solution. Blog:

    Prikaži ovu nit
    Poništi
  2. prije 15 sati

    I'm a big fan of the idea of smart glasses, and it looks like the technology is coming along. Even the prototype here looks pretty good. The part about not needing to focus since the image is painted directly on your retina is interesting and weird!

    Poništi
  3. 4. velj

    Meena is untested, but other language models have been revolutionary. Translation, natural language search, voice search, voice recognition, automated captioning, and many other tasks are all built on language models. Sign me up for the cargo cult, I guess? 10/10 ?

    Prikaži ovu nit
    Poništi
  4. 4. velj

    Communication: Narrow tasks like providing information to customers or even calling to make appointments for you are becoming the domain of machines. New knowledge: many linguists now see language models as useful tools for studying human language. 9/10

    Prikaži ovu nit
    Poništi
  5. 4. velj

    Twist: Every example relates to how deep learning chatbots are influencing our world. Passing impassable: even short range context awareness in dialogue was impossible 10 years ago. Looking down: chatbots are mapping and influencing our social landscapes, for good and ill 8/10

    Prikaži ovu nit
    Poništi
  6. 4. velj

    I'm not saying that Meena or models like it are as important as the balloon. But "statistical language models" are undoubtedly a small step on the path towards language understanding in . 7/10

    Prikaži ovu nit
    Poništi
  7. 4. velj

    Now, no-one would suggest that the balloon "solved long distance travel". But neither do planes: you can't take a plane into space! These are all small steps on the pathway of being able to reach into the unknown that may never be completed. 6/10

    Prikaži ovu nit
    Poništi
  8. 4. velj

    Balloons also allowed scientists to investigate the sky, learning a great deal about it that would go on to assist in the development of other aeronautical technologies, including the plane. 5/10

    Prikaži ovu nit
    Poništi
  9. 4. velj

    In reality, the electrical telegraph (which was invented around the same time) was a better solution, but if not for this it is likely balloon networks world have revolutionised long distance communication, being able to send signals over long distances than semaphore towers 4/10

    Prikaži ovu nit
    Poništi
  10. 4. velj

    They can take you higher than ever before, giving a broader view of the world. This revolutionises mapping (and unfortunately warfare as well, there are always harms with new tech). Perhaps this even allows us to look at how humans and societies work from a new perspective 3/10

    Prikaži ovu nit
    Poništi
  11. 4. velj

    It can take you short distances over otherwise impassable terrain, which was impossible before. This allows small groups of people to explore new regions. They can't actually make use of those regions as if they had roads, but it is still a worthwhile advance. 2/10

    Prikaži ovu nit
    Poništi
  12. 4. velj

    I like the flight analogy for models like , but draw a very different conclusion. Imagine humans can't fly, and someone invented a hot air balloon. Now, you haven't solved long distance travel, but this simple tool can still do a lot. 1/10

    Prikaži ovu nit
    Poništi
  13. proslijedio/la je Tweet
    4. velj

    Great opportunity if you're interested in radiology ML.

    Poništi
  14. 4. velj

    Am I missing something? I'm getting all these comments in my feed from big name researchers criticising claims that has "mastered language" or similar. Has someone said this? I'm assuming they are all referencing Meena, but the Google blog doesn't say anything like that.

    Poništi
  15. proslijedio/la je Tweet
    3. velj
    Odgovor korisnicima

    Just when I think I’m done with Twitter I see a gif like this

    Poništi
  16. 3. velj

    (obviously no slight intended in the first branch towards qualitative researchers - just assume I meant quantitative data)

    Prikaži ovu nit
    Poništi
  17. 3. velj

    Very thorough post on the *super* tired debate re: terminology ML vs statistics (usually seen in the context of yet another paper claiming that ML is silly because logistic regression performs as well on task x). My own contribution:

    Prikaži ovu nit
    Poništi
  18. 31. sij

    Really nice talk (and paper) by and team for an overview of causal problems and confounding in medical machine learning. I really like the framework approach (contextualising confounding by referencing the clinical process that causes it).

    Poništi
  19. 31. sij

    Nice sensible summary of Google's new "biggest language-ish model ever" from someone who builds chatbots.

    Poništi
  20. proslijedio/la je Tweet
    29. sij

    Great to see these developments- agencies will play an important role. Our forthcoming and international guidance for involving also very relevant!

    Poništi
  21. 30. sij

    Current scoring systems (SPAN, THRIVE, HIAT) are weak. W/ image and clinical data this pilot study reports improved AUC (0.75 vs 0.5-0.7). IMO this kind of work will have the greatest health impacts in . Extending rather than replacing or augmenting human effort.

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·